Security leaders live by the mantra that you can’t protect what you can’t see. For many organizations, the fear of shadow IT continues to plague their security teams as new apps are introduced to their stack. This visibility gap has only worsened with a flood of AI agents being introduced across apps or built in-house by their own teams.

As the foundation of our survey, we asked CISOs to rate the following statements from strongly agree to strongly disagree:

1. I know where my agents are

2. I know what my agents can connect to

3. I know what my agents can do

Less than half of CISOs feel confident they can identify all of the agents in their environment (47%), control what their agents interact with (46%), or authorize individual tool calls using context and intent clues (45%). The percentage drops below 30% if we only look at those who responded with “strongly agree.” This lack of confidence is pointing these leaders toward building or uncovering new solutions as quickly as possible.

France ranks last in oversight confidence. Despite high belief and anxiety that employees are using shadow AI (88% report at least some unsanctioned AI usage), French organizations rely on passive remediation: Nearly 24% only notify users and rely on policy compliance when shadow AI is found, making them uniquely exposed.

Interestingly, the research reveals a visibility vs. security paradox, where visibility alone doesn’t provide peace of mind. Even among CISOs who are fully confident they know where their AI agents are, roughly 80% remain highly concerned about excessive access and permissions. Knowing an agent's location or connections modestly tempers, but never eliminates, the worry of a breach.

The majority of security leaders aren’t feeling especially confident in securing their agents today, but not all companies (or regions) are at the same point in their journeys. Leaders were asked to rank their agentic identity and access governance by maturity:

• Reactive: “We are just beginning to discover AI usage and do not yet have formalized identity policies or controls for AI agents.”

• Developing: “We have basic policies in place for sanctioned AI, but lack comprehensive visibility and automated access controls.”

• Defined: “We have standardized identity controls (e.g., SSO, MFA) applied to known AI tools, but struggle with shadow AI or downstream agent permissions.”

• Advanced: “We have automated, real-time access controls, continuous monitoring, and dynamic governance for all human and AI identities.”

The UK is furthest along in its AI governance journey, with 36% of organizations reporting advanced, fully automated governance—the highest share of any country surveyed. That progress hasn't come from complacency: UK security leaders remain highly alert to risk, with 46% reporting they're "very concerned" about excessive AI access going unreviewed, second only to the US.

Mature identity governance offers protection from threats

While more mature organizations worry less about AI-enabled security breaches and reactive orgs are most likely to be extremely worried (32%), many of these reactive teams are making the trade-off to move fast. Nearly two-thirds (61%) of reactive organizations prioritize innovation over security, a 13-point increase over the average across all maturities (48%).

That said, there’s a clear path to protection: The more mature your AI identity governance becomes, the less you need to worry about breaches, shadow AI, and rogue agents that can’t be stopped.

Reactive companies report the most extensive shadow AI (25% compared with 13% across maturities) and are more likely to struggle to identify and stop rogue agents. Where advanced companies revoke access from rogue agents quickly (50% within minutes compared with 26% across the cohort), 18% of reactive companies can’t identify or stop them at all. Meanwhile, 55% of companies can revoke AI agent access within hours in the event of a breach—but even responding in hours is far too slow, as AI systems can execute harmful actions instantly.

Not a single company in the advanced tier struggles to identify and put a stop to these agents.

Gaining full control over AI identities is top priority

Security leaders are recognizing that the tools and processes built for human identity weren't designed for this moment—and they're moving quickly to catch up. To succeed in this new, complex environment, security teams are prioritizing governance as the path forward: putting identity at the center of their AI strategy so they can move to a fully automated, optimized, and secure agentic strategy. Maintaining full control over who has access to agents at any given moment was ranked as the most crucial AI security feature for securing the agentic enterprise by 29% of respondents, followed by agent discovery and credential management, each with 20%.

Japan reports the highest share of CISOs who are "extremely worried" about AI-driven breaches (29%)—more than any other market surveyed. They’re also twice as likely (12%) to report they have no consistent approach to governing agent identity compared with the global average (6%). Closing that gap will require board buy-in—but Japanese boards are among the most likely globally to treat AI security as a compliance hurdle rather than a business enabler, making it harder for CISOs to secure the investment they need.

This drive for better governance makes sense once you look at how AI identities are actually being managed today—the methods currently in place are, in many cases, still catching up to the risk. For example, one in four CISOs say they treat agents exactly as they do human identities. But current identity and security stacks were built for humans and traditional software that have predictable lifecycles and fixed execution paths. The non-deterministic nature of AI agents creates gaps that existing tools simply aren’t designed to close.

Even more worrisome, 21% of organizations use shared credentials or service accounts with broad permissions to govern AI access, and nearly the same amount (20%) leave agent management to the team that deployed them on an ad hoc basis.

How organizations manage AI agent access control

The security leaders working to mature their AI governance policies are moving as fast as they can, knowing that any unmanaged entities pose a threat to their organizations. This fear of AI-driven breaches is especially acute among US-based CISOs, with 84% of respondents feeling extremely or very worried (compared with 57% of CISOs globally).

Outside of breaches, 81% of global security leaders are deeply concerned about excessive AI access not being properly reviewed.

CISOs in Germany report the highest confidence in agent oversight but have some of the lowest actual governance maturity (58% developing or reactive). They’re also very concerned about AI phishing at nearly 81%, the highest single-threat signal in the study.

Digging into what's fueling this anxiety, security leaders point to a few specific threats. They cited AI-powered phishing attempts as the most concerning AI threat (61%), followed by malicious AI agents (49%), and deepfake authentication bypass (45%).

Reviewing the biggest barriers to agentic security

Across maturity levels and regional differences, security leaders found common ground in their biggest barriers to securing AI. A lack of visibility into AI identities is the most common barrier to security (48%), an especially prevalent challenge in the US market where 74% of respondents struggle with it. 

Other factors standing in the way of AI security include budget and headcount restraints (39%) and silos between AI and security teams (33%), which make it difficult to implement the necessary changes. Canadian CISOs stood out from the group, believing their biggest barrier to securing the agentic enterprise lies in a lack of an industry standard (65%).

Biggest barriers to securing AI with identity-centric controls

Top AI security barrier by country (#1 reported challenge)

Bringing shadow AI into the light

Over two-thirds (68%) of CISOs see at least some unsanctioned use of AI and agentic AI. With the limitations outlined above, teams are doing what they can to manage and respond to shadow AI threats wherever possible. The most common response overall when shadow AI is detected is to block access (35%) or bring the tool under governance and apply access controls (29%) with whatever solutions the team has available.

Executive leaders may recognize that identity is core to AI security, but there’s still a substantial gap between recognition of a strategy and complete alignment on its execution. Less than a third (31%) of CISOs feel they’re fully aligned with their C-suite and board on acceptable levels of AI-related risk. US-based leaders feel the least aligned of the cohort, with only 12% stating they feel fully aligned, an unsurprising finding given their outsized concern about breaches.

US organizations show the highest breach anxiety (84%) and high concern over overprivileged agents (65%), but they’re better positioned for success because a majority of their boards (54%) see security as a business enabler.

CISOs have been hard at work convincing their boards and C-suite executives that security is a business enabler rather than an innovation blocker, but in the age of AI, it’s become even more difficult to make that case. Less than half of CISOs (46%) feel their boards view AI security as a business enabler today.

However, this friction varies significantly by region. Boards in the US and France are the most likely to view AI security as a business enabler, making them more open to investment. German and Japanese organizations, however, are the most likely to view AI security merely as a compliance or regulatory hurdle. This lack of executive understanding creates a top-down disconnect that leaves AI identity security under-resourced.

It’s up to CISOs to make it clear they’re a business team player. This advocacy may be the deciding factor between a secure agentic enterprise and one that’s stuck in a reactive state.